Universe Today
Monitoring Earth from above is a messy, inconsistent business. Clouds can block a satellite view, the angle of the Sun can create huge shadows, and orbital paths might not take a system across the same general area for weeks. That inconsistency is hard to deal with for traditional software, and can even confuse standard deep learning algorithms. But, according to a new review paper, available in pre-print by researchers Raul-Alexandru Gorgan and Dorian Gorgan of the Technical University of Cluj-Napoca, there’s a new type of frontier AI that could be able to help - bio-inspired Liquid Neural Networks.
Modern society relies heavily on Earth observation, but collecting data in a meaningful way that is understandable to a deep learning system is complicated. For example, most deep learning networks, such as those that drive autonomous cars, expect data points to come in at regular intervals, and usually for those intervals to be relatively short, so that small changes can be distinguished and used to inform the model.
That level of regularity doesn’t happen in Earth observation. The most obvious culprit is clouds. Satellites in Low Earth Orbit (LEO) can have their views of certain areas over the planet blocked by clouds - repeatedly in many cases. Combine that with relatively erratic flyover schedules, and some satellites can go over a month between being able to take pictures of certain parts of the world.
Fraser talks about whether Earth observation satellites are watching you from space.A lot can change in a month, both through gradual processes, such as fields growing over a summer, but also through catastrophic one-off events, like flash floods or volcanoes. Traditional processing methods, such as Convolutional Neural Networks (CNNs) or Visual Transformers, while great for processing time in small, countable steps, do not do well with filling in large gaps where a lot has happened in between . In other words, they are confused by the crop growth, or the distinct wildfire boundary.
There are plenty of solutions on the table to tackle this problem, and the review paper takes a look at 61 different studies published between 2018 and 2026, showing how artificial intelligence is evolving in response to these challenges. Many adaptive techniques use extensions of today's traditional computer vision models, adding graph-based models or dynamic receptive fields to track changes in building footprints or dehaze foggy areas.
But perhaps the most interesting approach is the Liquid Neural Networks. These are inspired by the simple, compact nervous systems of C. Elegans, a microscopic roundworm commonly used in biological studies, which only contains 302 total neurons. Unlike traditional software models, LNNs use a type of calculus equation known as an Ordinary Differential Equation (ODE) to allow it to adapt to a continuous flow of time even when there are huge chunks of data missing.
Fraser and Pamela discuss remote sensing and how it's so important to life as we know it.If a satellite happens to miss three consecutive passes over a farm due to cloud cover, the LNN can easily integrate that time gap mathematically, making it so it seems like the system never missed the data points. They also adapt to dramatic changes as well, and are confused by significant events like volcanic eruptions or wildfires. And since they use ODE, which aren’t as computationally intensive as brute force code, they could eventually run on smaller, lower-power computer chips - potentially even directly on some of the satellites that are doing the observing.
It’s still early days for the technology, though. Of the 61 studies in the review article, only around 16% of them used LNNs. But even more critically, many of the studies only tested their code in a single region, or didn’t take note of how much computing power was used to complete their calculations. In other words, researchers should temper their enthusiasm about this new technology until it can really prove it offers a distinct advantage to these observational problems.
Soon it will have plenty of opportunities to prove itself. Space agencies and private companies alike keep launching new Earth-observation satellites on a regular basis, providing yet more training data. As researchers work through what technique works best, our understanding of how our own planet works will continue to get better.
Learn More:
R.-A. Gorgan & D. Goran - Adaptive Neural Networks for Remote Sensing Imagery: A Systematic Review
UT - A Unique Combination of Antennas Could Revolutionize Remote Sensing
UT - Earth Observation Satellites Could be Flown Much Lower than Current Altitudes and Do Better Science
UT - Earth is an Exoplanet to Aliens. This is What They'd See
